Defensive Refusal: A Qualitative Framework for Escalatory Rejection Behavior in Large Language Models

This paper presents a qualitative observational framework for a failure mode in large language models termed "Defensive Refusal," characterized by progressive escalation toward refusal in response to user correction rather than epistemic revision. The study adopts an event-triggered observational methodology, presenting one complete illustrative interaction log alongside abstracted cross-domain observations under responsible disclosure constraints. The work is hypothesis-generating and does not claim statistical prevalence. This manuscript is currently under moderation at arXiv. The Zenodo version serves as a citable, archived reference of the present work.